Federal relationship perception learning system, method and device based on financial knowledge graph

Abstract

The application discloses a federal relationship-aware learning system, method and device based on an alternating early fusion technology. The system comprises a central server and a plurality of clients. Each client is composed of a federal client and a local model. The local model and entity embedding of each client are unique. The system has privacy, universality and high efficiency in the case that different mode entity data and different local models exist in each client.

Description

Technical Field

[0001] This application relates to the field of learning systems, and in particular to federated relationship-aware learning systems, methods, and apparatus based on financial knowledge graphs. Background Technology

[0002] A knowledge graph, also known as a semantic network, is a network that records real-world financial entities and their relationships. While graph-based methods have been widely used to represent knowledge in the past, financial knowledge graphs offer a powerful approach that combines artificial intelligence with graph computing to manage complex financial data. Many financial institutions are interested in building financial knowledge graphs, but the cost and effort involved in doing so are often very high. It would be beneficial if two or more financial institutions could share their data, allowing each institution to leverage data disclosed by other institutions to complete its own financial knowledge graph. However, privacy regulations such as GDPR (General Data Protection Regulation) or CCPA (California Consumer Privacy Act) severely hinder data-sharing practices: for example, in the financial or healthcare industries, most of the data collected is private and cannot be disclosed to others. Under privacy regulations like GDPR and CCPA, holders of financial knowledge graphs must keep their private data inaccessible.

[0003] To break down data silos, Google proposed Federated Learning, which can train a global model using local model updates without collecting private data. FedAvg is a representative algorithm of Federated Learning, its core being to update the global model by uploading and aggregating locally generated updates (such as gradients or model parameters) from clients to a central server, and then updating the global model through a weighted average. Building on FedAvg, some works have attempted to combine financial knowledge graph completion tasks with Federated Learning, but most have only performed simple combinations without further leveraging the properties of financial knowledge graph completion tasks within a federated learning environment. For example, Figure 1 This paper demonstrates a simple framework for performing a financial knowledge graph completion task in a federated learning environment. Three institutions update their local financial entity embeddings into their financial knowledge graph completion model and send them to a server. The server receives all the embedded local financial entities and computes the embedded global financial entities using FedAvg or another federated learning method. The server then sends the global financial entities embedded into all institutions. Each institution receives its embedded global financial entities and begins a new round of training until a stopping condition is validated.

[0004] However, the aforementioned framework is insufficient to meet the requirements. In some model environments, servers and institutions not only exchange embedded financial entities but also model weights. Servers must maintain a large number of financial entity embedding matrices while simultaneously updating global model weights and embeddings, a process that consumes significant memory and processing resources. Furthermore, clients or servers can discover private information by combining model updates with financial entity embeddings, leading to reduced security for this private information.

[0005] Therefore, the aforementioned technical problems in the relevant technologies urgently need to be solved. Summary of the Invention

[0006] This application aims to address one of the technical problems in related technologies. To this end, embodiments of this application provide a federated relationship-aware learning system, method, and apparatus based on financial knowledge graphs, which can ensure the security of private financial information while improving the model's versatility.

[0007] According to one aspect of the embodiments of this application, a federated relationship-aware learning system based on financial knowledge graph is provided. The system includes: a central server and several financial institution clients. Each client consists of a federated client and a local model. The local model and financial entity embedding of each client are unique.

[0008] The central server receives local updates and local financial relationship embeddings from the client. The global model is trained in a cross-domain parallel adaptive manner and alternately fuses the embedded relationship-aware federated learning part and the local model during the training process. It uses financial relationship data to perform federated aggregation and alternation to complete the financial knowledge graph completion and distributes the global model weights and global financial relationships to each client.

[0009] The federated client receives global financial relationship embeddings from the central server, utilizes alternating fusion, uses local financial entities embedded in ent_i to improve the global model and global financial relationship embeddings, updates the financial relationship embeddings through contextual relationships and relationship paths, and sends the locally updated weights and financial relationship embeddings into rel_i to the central server; the local model receives global financial relationship embeddings from the central server, utilizes federated aggregation and alternation to achieve highly sensitive and adaptable utilization and enhancement of additional local financial entity data and global data, and uses g_rel to fuse rel_i to enhance the local model.

[0010] In one embodiment, the local model includes a financial knowledge graph completion model, which has unified relation information and a preset number of relation embeddings.

[0011] In one embodiment, the federated client consists of a relationship context and a relationship path;

[0012] The financial relationship context is used to describe the proximity relationship between head financial entities and tail financial entities, and to obtain information on the adjacency relationship of edges.

[0013] The financial relationship path is used to determine the relative positions of head and tail financial entities.

[0014] In one embodiment, the formula for calculating the financial relationship context is:

[0015]

[0016] In the formula, c(h,t) represents the financial relationship context, σ is the activation function, h represents the head financial entity, and t represents the tail financial entity. The model weights that the client can learn during the training of the (k-1)th model. This represents the hidden state of the top financial entities in the client during the (k-1)th iteration. This represents the hidden state of the tail financial entities in the client during the (k-1)th iteration.

[0017] In one embodiment, the formula for calculating the financial relationship path is:

[0018]

[0019] In the formula, Let h be the head financial entity and t be the tail financial entity. and Let p be the learnable relational path weights for the client, and p be the set of paths between head-end and tail-end financial entities, and each path in the set. Let h be the set of paths connecting the head h and the tail t, where mean is the average function and linear is the linear function.

[0020] In one embodiment, the federated client obtains global weights and global financial relationship embeddings based on the financial relationship context and financial relationship path, and the calculation method includes:

[0021]

[0022] In the formula, For the global model weights of the client, To update model weights locally on the client side, The total number of triples across all clients. For all financial entities in the i-th client.

[0023] In one embodiment, the system embeds a local model by combining the global financial relationship and the local financial relationship, as shown in the formula:

[0024]

[0025] In the formula, rel_i is the local financial relationship embedding in the i-th client, g_rel is the global financial relationship embedding in the federated learning framework, and mean is the average value function.

[0026] According to one aspect of the embodiments of this application, a federated relationship-aware learning method based on financial knowledge graphs is provided, characterized in that it is applied to the federated relationship-aware learning system based on financial knowledge graphs described in the preceding embodiments, and the method includes:

[0027] The global model receives local updates and local financial relationship embeddings from the client. It is trained in a cross-domain parallel adaptive manner and alternately fuses the embedded relationship-aware federated learning part and the local model during the training process. It uses financial relationship data to perform federated aggregation and alternation to complete the financial knowledge graph. The global model weights and global financial relationships are then distributed to each client.

[0028] Receive global financial relationship embeddings from the central server, utilize alternating fusion, use local financial entities embedded in ent_i to improve the global model and global financial relationship embeddings, update the financial relationship embeddings through contextual relationships and relationship paths, and send the locally updated weights and financial relationship embeddings into rel_i to the central server.

[0029] Global financial relationship embeddings are received from the central server, and federated aggregation and early alternation are used to achieve highly sensitive and flexible adaptive enhancement of additional local financial entity data and global data. g_rel is used to fuse rel_i to enhance the local model.

[0030] According to one aspect of the embodiments of this application, a federated relationship-aware learning device based on a financial knowledge graph is provided, characterized in that the device comprises:

[0031] The first module is used to receive local updates and local financial relationship embeddings from the client. The global model is trained in a cross-domain parallel adaptive manner and alternately fuses the embedded relationship-aware federated learning part and the local model during the training process. It uses financial relationship data to perform federated aggregation and alternation to complete the financial knowledge graph completion and distributes the global model weights and global financial relationships to each client.

[0032] The second module is used to receive global financial relationship embeddings from the central server, use alternating fusion, use local financial entities embedded in ent_i to improve the global model and global financial relationship embeddings, update the financial relationship embeddings through contextual relationships and relationship paths, and send the locally updated weights and financial relationship embeddings into rel_i to the central server.

[0033] The third module is used to receive global financial relationship embeddings from the central server, and leverages federated aggregation and early alternation to achieve highly sensitive and flexible adaptive enhancement of additional local financial entity data and global data. It uses g_rel to fuse rel_i to enhance the local model.

[0034] According to one aspect of an embodiment of this application, a federated relationship-aware learning device based on a financial knowledge graph is provided, the device comprising:

[0035] At least one processor;

[0036] At least one memory for storing at least one program;

[0037] When at least one of the programs is executed by at least one of the processors, it implements the federated relationship-aware learning system or the federated relationship-aware learning method based on financial knowledge graphs as described in the preceding embodiments.

[0038] The beneficial effects of the federated relation-aware learning system, method, and apparatus based on financial knowledge graph provided in this application are as follows: The system of this application includes a central server and several clients. Each client consists of a federated client and a local model. The local model and financial entity embedding of each client are unique. The central server receives local updates and local relation embeddings from the clients. The global model is trained in a cross-domain parallel adaptive manner and alternately fuses the embedded relation-aware federated learning part and the local model during the training process. It uses relation data to perform federated aggregation and early alternation to complete the KG, and distributes the global model weights and the global relations to each client. The federated client receives global relation embeddings from the central server, uses alternating fusion, uses the local financial entities embedded in ent_i to improve the global model and global relation embeddings, updates the relation embeddings through context relations and relation paths, and sends the locally updated weights and relation embeddings into rel_i to the central server. The local model receives global relation embeddings from the central server, uses federated aggregation and early alternation to achieve ultra-sensitive utilization and plastic adaptive enhancement of additional local financial entity data and global data, and uses g_rel to fuse rel_i to enhance the local model. The system proposed in this application is capable of providing privacy, versatility, and efficiency, even with different financial entity data patterns and different local models that may exist on each client.

[0039] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 A schematic diagram of a federation relationship awareness learning system based on financial knowledge graphs provided in an embodiment of this application;

[0042] Figure 2 A flowchart of a federated relationship-aware learning method based on financial knowledge graphs provided for embodiments of this application;

[0043] Figure 3 A schematic diagram of a federation relationship awareness learning device based on a financial knowledge graph provided in an embodiment of this application;

[0044] Figure 4 A schematic diagram of another federation relationship awareness learning device based on financial knowledge graph provided for an embodiment of this application. Detailed Implementation

[0045] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0046] The terms "first," "second," "third," and "fourth," etc., used in the specification, claims, and drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0047] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0048] A knowledge graph (KG), also known as a semantic network, is a network that records real-world financial entities (such as people, things, places, historical events, etc.) and their relationships, referred to as financial entities (or concepts) and their relationships. While graph-based methods have been widely used to represent knowledge in the past, knowledge graphs offer a powerful approach that combines artificial intelligence methods with graphs to manage complex data. Many companies are interested in building KGs, but the cost and effort involved in building and completing them are often very high. Therefore, a constructed KG is often not very useful because it ignores many financial entities and relationships. It would be beneficial if two or more companies could share their data, and in this way, each company could utilize data disclosed by other companies to complete its own KG. However, privacy regulations such as GDPR (General Data Protection Regulation) or CCPA (California Consumer Privacy Act) strictly hinder data-sharing practices: for example, in the financial or healthcare industries, most of the data collected is private and cannot be disclosed to others. Under privacy regulations such as GDPR and CCPA, graph holders must keep private data inaccessible.

[0049] To break down data silos, Google proposed Federated Learning, which can train a global model using local model updates without collecting private data. FedAv is a representative algorithm of Federated Learning, its core being to update the global model by uploading and aggregating locally generated updates (such as gradients or model parameters) from clients to a central server, and then updating the global model through a weighted average. Building on FedAv, some works have attempted to combine financial knowledge graph completion (knowledge graph completion) tasks with Federated Learning, but most have only performed simple combinations without further leveraging the properties of financial knowledge graph completion tasks in a federated learning environment. For example, Figure 1This paper demonstrates a simple framework for performing a financial knowledge graph completion task in a federated learning environment. Three institutions update their local financial entity embeddings into their financial knowledge graph completion model and send them to a server. The server receives all the embedded local financial entities and computes the embedded global financial entities using FedAvg or another federated learning method. The server then sends the global financial entities embedded into all institutions. Each institution receives its embedded global financial entities and begins a new round of training until a stopping condition is validated.

[0050] However, the aforementioned framework is insufficient to meet the requirements. In some model environments, servers and institutions not only exchange embedded financial entities but also model weights. Servers must maintain a large number of financial entity embedding matrices while simultaneously updating global model weights and embeddings, a process that consumes significant memory and processing resources. Furthermore, clients or servers can discover private information by combining model updates with financial entity embeddings, leading to reduced security for this private information.

[0051] To address the aforementioned issues, this application proposes a federated relationship-aware learning system, method, and apparatus based on financial knowledge graphs. This application presents a general and efficient federated knowledge graph completion framework capable of collaboratively training a high-quality financial knowledge graph completion model using graph data distributed across multiple institutional clients, while simultaneously protecting data privacy. In the model of this application, each institution has a unique association with a client. Any i-th financial institution client maintains a financial knowledge graph Gi=(Vi,Ei), where Vi contains the financial entities corresponding to the i-th client (such as corporate accounts, credit products, customer IDs, etc.), and Ei... Vi×Vi is an edge set. Each edge identifies the relationship between the financial entities it connects (such as lending, guarantee, investment, etc.). Each client attempts to predict the financial relationships between nodes in Gi. Given the realities, it's conceivable that even within the same domain, the number of nodes, graph structure, and data distribution can vary slightly between different clients. More importantly, any client may contain multiple forms of data, such as images, text, etc., which we call additional information. Furthermore, these institutions often share the same type of relationship between a pair of financial entities.

[0052] like Figure 1The diagram illustrates the federated relation-aware learning system based on financial knowledge graphs proposed in this application. It includes a central server and several clients. Each client consists of a federated client and a local model. Each client's local model and financial entity embeddings are unique. The central server receives local updates and local relation embeddings from the clients. The global model is trained in a cross-domain parallel adaptive manner, alternately fusing the embedded relation-aware federated learning part and the local model during training. Relation data is used for federated aggregation and early / late alternation to complete the knowledge graph (KG). The global model weights and the global relations are distributed to each client. Specifically, the central server receives local updates and local relation embeddings from the clients. The global model is trained in a cross-domain parallel adaptive manner, alternately fusing the embedded relation-aware federated learning part and the local model during training. Relation data is used for federated aggregation and early / late alternation to complete the KG. The global model weights and the global relations are distributed to each client. The global model weights and global relationships are distributed to each client. Each federated client receives the global relationship embedding from the central server, utilizes Alternating fusion (AEF), uses local financial entities embedded in ent_i to improve the global model and global relationship embedding, updates the relationship embedding through contextual relationships and relationship paths, and sends the locally updated weights and relationship embeddings into rel_i to the central server. The local model receives the global relationship embedding from the central server, utilizes federated aggregation and early alternation to achieve highly sensitive and adaptable enhancement of additional local financial entity data and global data, and uses g_rel to fuse rel_i to enhance the local model.

[0053] In this embodiment, the local model and financial entity embedding are unique for each client. For example, client 1 might have a local model TransE, while client 2 might have a local model TransD. Within each client, both the federated client and the local model can utilize Alternating Early Fusion (AEF). For instance, loc_i can use g_rel to fuse rel_i to enhance the local model. Simultaneously, fed_i can use the local financial entity embedded in ent_i to improve the global model and global relation embedding.

[0054] The local model includes a financial knowledge graph completion model, which has unified relation information and a preset number of relation embeddings.

[0055] The core processing components of this application's system are federal relationship-aware financial knowledge graph completion learning and alternating early fusion, specifically:

[0056] (1) Learning to complete the knowledge graph of federal relations-aware finance

[0057] Existing federal financial knowledge graph completion frameworks cannot support the needs of different clients to possess data in different ways. Because they ignore financial entity information and rely solely on topological information, the central server is forced to maintain a massive financial entity embedding matrix for all clients, creating a significant burden. Therefore, this application can efficiently complete the KG using only relational information, and in a federated learning scenario, KGs associated with different institutions within the same department share the same type of relations.

[0058] The federated client consists of relation contexts and relation paths. The relation contexts describe the proximity relationships between head and tail financial entities, obtaining information about the adjacency of edges. The relation paths determine the relative positions of head and tail financial entities. For example, given a triple (h, r, t), relation contexts ch and ct can collect information about the neighbors of edges of type r. Taking a bank as an example, if r is a "loan project," the surrounding relationships of h are "customer credit rating," "corporate entity," etc., and the surrounding relationships of t are "loan project type," "application," "cancellation," etc. Therefore, when identifying relation type r, the context of the relationship between h and t will be additional information.

[0059] Specifically, the formula for calculating the relational context is:

[0060]

[0061] In the formula, For the context of the relationship, Let h be the activation function, and t be the head financial entity and t be the tail financial entity. The model weights that the client can learn during the training of the (k-1)th model. This represents the hidden state of the top financial entities in the client during the (k-1)th iteration. This represents the hidden state of the tail financial entities in the client during the (k-1)th iteration.

[0062] Specifically, for each client c i ,c jie c represents the hidden state of edge e during the j-th iteration. jiv This provides relational context information about the financial entity in the j-th iteration. N(v) and N(e) represent the set of adjacent edges of the financial entity and the set of adjacent nodes of edge e, respectively. iv and cie The computation method is inspired by the classic node-based graph messaging mechanism. Therefore, after repeating the process more than k-1 times, the relation context pair c(h,t) can be computed. In this federated configuration, edge c 0ie The initial features are initialized by the relation type embedding rel_i, which is randomly initialized.

[0063] Relational context helps to obtain information about the adjacency relationships of edges, but it cannot obtain the relative positions of h and t. Therefore, the relational path between h and t helps to determine information about the relative positions of h and t.

[0064] Specifically, the formula for calculating the relationship path is:

[0065]

[0066] In the formula, Let h be the head financial entity and t be the tail financial entity, representing the relationship path. Let p be the learnable relational path weights for the client, and p be the set of paths between head-end and tail-end financial entities, and each path in the set. Let h be the set of paths connecting the head h and the tail t. Find the average function. It is a linear function.

[0067] Here, P(h,t) is the set of paths connecting the head h and the tail t, where each path p={r eo ,r e1 ,...,r ei}∈P(h,t) is formed by each edge e in the path k Related relation type r ek Composition, each relation r ek Represented by a single embedding vector. Here, `linear` is a function that performs a linear transformation. An aggregate representation of the relation path is constructed by averaging all paths from h to t. We can predict the missing relation between (h, t) by combining the relation context and the relation path, and using the softmax function, as shown in Equation 3. For training, we minimize the cross-entropy loss J between the predicted and true values ​​on the training triples Di, and we update each client c. i Local parameters.

[0068]

[0069] The global weight W can be updated via the central server. gAt the same time, importance is assigned to the local model weights from each client, thereby reducing the impact of data imbalance.

[0070] The federated client performs the above calculations on its own, and then uses the federated method FedAvg to update the global weight Wg for all clients on the central server. Finally, we can obtain the global weight using the formula. The global relationship is embedded in g_rel.

[0071] From the process of learning to complete the financial knowledge graph based on federated relationships, we can obtain a global relationship embedding g_rel, which can learn some topology and features of all clients. Since the data is distributed across local datasets on the clients, the local relationship embedding rel_i may not learn enough information during local model training. The federated clients obtain global weights and global relationship embeddings based on the relationship context and relationship path. The calculation method includes:

[0072]

[0073] In the formula, Indicates client c i The number of triples in the middle, and This represents the total number of triples across all clients. For the global model weights of the client, To update model weights locally on the client side, The total number of triples across all clients. For all financial entities in the i-th client.

[0074] (2) Alternating early fusion

[0075] Federated relation-aware financial knowledge graph completion learning can generate a global model even without financial entity information, addressing Problem 1 in Section 2.1. It is noteworthy that during the learning process, the federated clients only consider the relational information contained across all clients. However, since different clients typically have different data patterns and distributions, the global model may perform poorly on local datasets. Clients train their financial knowledge graph completion models on local data using various models or data processing methods; this application refers to these as local models. For example, client 1 might use TransE, while client 2 might use TransR. Many studies have proposed fusing multimodal features in multimodal representation learning. Based on this, this application proposes alternating early fusion to better utilize the additional information available to clients and meet the scenarios of different clients. Therefore, this application attempts to combine the financial entity embeddings of each client with their relation embeddings.

[0076] Suppose the client has financial entity information, such as text information, and the client uses a local model (such as TransE) to learn the local financial entity embedding ent_v. Then, as shown in Equation 5, ent_v can be used as additional information to assist in updating the relation embedding:

[0077] The system combines the global relationship and the local relationship to embed a local model, as shown in the formula:

[0078]

[0079] In the formula, Let V be the hidden state of financial entity v in client i during the j-th iteration. Let N(v) be the hidden state of edge e in client i during the j-th iteration, N(v) be the set of adjacent edges of v, and ent_v be the local financial entity embedding in client v.

[0080] From the process of learning to complete the financial knowledge graph with a federated relationship awareness, a global relationship embedding g_rel can be obtained, which can learn some topology and features of all clients. Since the data is distributed across local datasets on the clients, the local relationship embedding rel_i may not learn enough information during local model training. Therefore, this application combines the global relationship embedding and the local relationship embedding to improve the local relationship embedding, as shown below:

[0081]

[0082] In the formula, Local relation embedding in the i-th client, This is the global relation embedding in the federated learning framework, where mean is the average value function.

[0083] To further understand early alternating fusion, it's important to note that the client enhances the global model through federated relationship-aware learning and local financial entity embeddings, which can aid in updates. Conversely, `g_rel` can also be used to enhance the local relationship embedding `rel_i`, thereby improving the performance of the local model. These two processes can be switched independently to improve the overall framework performance. The alternating early fusion approach can be flexibly adjusted based on different data conditions. As an example of a training scenario, once a round of federated relationship-aware training is complete, the new global relationship embedding can be used to enhance the embedding learning of the local model `TransE`. The updated embedding of `TransE` can then be used for the next federated learning training, thus improving the performance of both the local and federated models through continuous alternating fusion.

[0084] To verify the technical effectiveness and feasibility of this application, experiments were conducted on four datasets, as shown in Table 1:

[0085] Table 1: Experimental Dataset Table

[0086]

[0087] In the table above, #Ent, #Rel, #Train, and #Test represent the number of financial entities, relations, training triples, and test triples, respectively. Among them, FB15K-237

[19] is a variant of the original Freebase dataset FB15k that does not include inverse relations; WN18RR[8] is a dataset created from WN18, a subset of WordNet, which removes inverse relations; NELL995

[27] is a dataset built from the 995th iteration of the NELL system; and DDB14

[23] is a dataset derived from a disease database that contains terms related to diseases, drugs, and their interactions.

[0088] The TransE model, with its transitional features, is currently the most widely used financial knowledge graph completion model. Each relation is defined by RotatE as a rotation from the head to the tail of the financial entity in a complex vector space. DistMult is a popular tensor factorization-based model that uses a bilinear fractional function to compute the score of knowledge triples. ComplEx is a DistMult extension that embeds financial entities and relations into complex vectors instead of real-valued vectors.

[0089] Based on this, this experiment compares the performance of the global model on the global test dataset and compares the performance of the global model on the local test dataset.

[0090] For a given pair of financial entities (h, t) in the test set, the underlying truth relation type r is sorted against all other candidate relation types. Pathcon changed the negative sampling method, transforming the relation r of each true triple (h, r, t) into relation r', and using (h, r', t) as a negative sample; here, r' comes from the set of relation types, while FedE sorts the true tail financial entity t against all other candidate tails for a given pair (h, r) in the test set. The financial entity numbers in each client are shown in Table 2.

[0091] Table 2: Financial Entity Numbers in Each Client

[0092]

[0093] F, W, N, and D represent datasets FB15K237, WN18RR, NELL995, and DDB14, respectively. * indicates that the datasets are split in an imbalanced setting, and overlapped indicates the number of overlapping financial entities across the five clients.

[0094] Table 3: Relationship prediction results for all datasets

[0095]

[0096] As shown in Table 3, the experimental results of this application are as follows: Distmult and ComplEx perform poorly in the federated scenario; for each dataset, their performance is significantly worse than the complete training set. TransE and RotatE do not cause significant performance loss in the federated learning environment, but their results are inferior to FedR financial knowledge graph completion. Equation 4 shows that the more data a client has, the greater its impact on the global model. The global test dataset is similar to client 1's local dataset in the imbalanced setting; therefore, FedR financial knowledge graph completion performs better in the imbalanced setting than in the balanced setting. In conclusion, FedR financial knowledge graph completion performs well on all datasets in the federated setting.

[0097] Furthermore, this application proposes a federated relationship-aware learning method based on financial knowledge graphs, which is applied to a federated relationship-aware learning system based on financial knowledge graphs, such as... Figure 2 As shown, the method includes:

[0098] S201. Receive local updates and local relation embeddings from the client. The global model is trained in a cross-domain parallel adaptive manner and the embedded relation-aware federated learning part and the local model are alternately fused during the training process. The relation data is used for federated aggregation and alternation to complete the KG. The global model weights and global relations are distributed to each client.

[0099] S202. Receive the global relation embedding from the central server, use alternating fusion, use the local financial entity embedded in ent_i to improve the global model and global relation embedding, update the relation embedding through context relations and relation paths, and send the locally updated weights and relation embeddings into rel_i to the central server.

[0100] Furthermore, this application also proposes a federated relationship-aware learning device based on financial knowledge graphs, such as... Figure 3 As shown, the device includes:

[0101] The first module 301 is used to receive local updates and local relation embeddings from the client. The global model is trained in a cross-domain parallel adaptive manner and alternately fuses the embedded relation-aware federated learning part and the local model during the training process. The relation data is used to perform federated aggregation and alternate early and early to complete the KG. The global model weights and global relations are distributed to each client.

[0102] The second module 302 is used to receive global relation embeddings from the central server, improve the global model and global relation embeddings by using alternating fusion and local financial entities embedded in ent_i, update the relation embeddings through context relations and relation paths, and send the locally updated weights and relation embeddings into rel_i to the central server.

[0103] The third module 303 is used to receive global relation embeddings from the central server, utilize federated aggregation and early alternation to achieve highly sensitive and flexible adaptive enhancement of additional local financial entity data and global data, and use g_rel to fuse rel_i to enhance the local model.

[0104] Furthermore, this application also proposes a federated relationship-aware learning device based on financial knowledge graphs, such as... Figure 4 As shown, the device includes:

[0105] At least one processor;

[0106] At least one memory for storing at least one program;

[0107] When at least one of the programs is executed by at least one of the processors, it implements the federated relationship-aware learning system or the federated relationship-aware learning method based on financial knowledge graphs as described in the preceding embodiments.

[0108] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this application are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is changed and sub-operations described as part of a larger operation are executed independently.

[0109] Furthermore, although this application is described in the context of functional modules, it should be understood that, unless otherwise stated to the contrary, one or more of the functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding this application. Rather, given the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed herein, the actual implementation of the module will be understood within the scope of conventional technology for an engineer. Therefore, those skilled in the art can implement the application set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and not intended to limit the scope of this application, which is determined by the full scope of the appended claims and their equivalents.

[0110] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0111] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0112] In the foregoing description of this specification, the references to terms such as "one embodiment," "another embodiment," or "some embodiments," etc., indicate that a specific feature, structure, material, or characteristic described in connection with an embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0113] Although embodiments of this application have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the claims and their equivalents.

[0114] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A federated relationship-aware learning system based on financial knowledge graphs, characterized in that, The system includes a central server and several clients, each client consisting of a federated client and a local model, and each client's local model and financial entity embedding are unique; The central server receives local updates and local financial relationship embeddings from the client. The global model is trained in a cross-domain parallel adaptive manner and alternately fuses the embedded relationship-aware federated learning part and the local model during the training process. It uses financial relationship data to perform federated aggregation and alternation to complete the financial knowledge graph completion and distributes the global model weights and global financial relationships to each client. The federated client receives global financial relationship embeddings from the central server, utilizes alternating fusion, uses local financial entities embedded in ent_i to improve the global model and global financial relationship embeddings, updates the financial relationship embeddings through contextual relationships and relationship paths, and sends the locally updated weights and financial relationship embeddings into rel_i to the central server; the local model receives global financial relationship embeddings from the central server, utilizes federated aggregation and alternation to achieve highly sensitive and adaptable utilization and enhancement of additional local financial entity data and global data, and uses g_rel to fuse rel_i to enhance the local model.

2. The federated relationship-aware learning system based on financial knowledge graphs according to claim 1, characterized in that, The local model includes a financial knowledge graph completion KGC model, which has unified financial relationship information and a preset number of financial relationship embeddings.

3. The federated relationship-aware learning system based on financial knowledge graphs according to claim 1, characterized in that, The federal client consists of a financial relationship context and a financial relationship path; The financial relationship context is used to describe the proximity relationship between head financial entities and tail financial entities, and to obtain information on the adjacency relationship of edges. The financial relationship path is used to determine the relative positions of head and tail financial entities.

4. The federated relationship-aware learning system based on financial knowledge graphs according to claim 3, characterized in that, The formula for calculating the financial relationship context is as follows: In the formula, c(h,t) represents the financial relationship context, σ is the activation function, h represents the head financial entity, and t represents the tail financial entity. The model weights that the client can learn during the training of the (k-1)th model. This represents the hidden state of the top financial entities in the client during the (k-1)th iteration. This represents the hidden state of the tail financial entities in the client during the (k-1)th iteration.

5. The federated relationship-aware learning system based on financial knowledge graphs according to claim 3, characterized in that, The formula for calculating the financial relationship path is as follows: In the formula, Let h be the head financial entity and t be the tail financial entity. and Let p be the learnable relational path weights for the client, and p be the set of paths between head-end and tail-end financial entities, and each path in the set. Let h be the set of paths connecting the head h and the tail t, where mean is the average function and linear is the linear function.

6. The federated relationship-aware learning system based on financial knowledge graphs according to claim 3, characterized in that, The federated client obtains global weights and global financial relationship embeddings based on the financial relationship context and financial relationship path, and the calculation method includes: In the formula, For the global model weights of the client, To update model weights locally on the client side, The total number of triples across all clients. For all financial entities in the i-th client, Indicates client The number of triples in the middle This represents the i-th client.

7. The federated relationship-aware learning system based on financial knowledge graphs according to claim 1, characterized in that, The system combines the global financial relationships and the local financial relationships into a local model, as shown in the formula: In the formula, rel_i is the local financial relationship embedding in the i-th client, g_rel is the global financial relationship embedding in the federated learning framework, and mean is the average value function.

8. A federated relationship-aware learning method based on financial knowledge graphs, characterized in that, The method, applied to the federated relationship-aware learning system based on financial knowledge graphs as described in claim 1, comprises: The global model receives local updates and local financial relationship embeddings from the client. It is trained in a cross-domain parallel adaptive manner and alternately fuses the embedded relationship-aware federated learning part and the local model during the training process. It uses financial relationship data to perform federated aggregation and alternation to complete the financial knowledge graph. The global model weights and global financial relationships are then distributed to each client. Receive global financial relationship embeddings from the central server, utilize alternating fusion, use local financial entities embedded in ent_i to improve the global model and global financial relationship embeddings, update the financial relationship embeddings through contextual relationships and relationship paths, and send the locally updated weights and financial relationship embeddings into rel_i to the central server. Global financial relationship embeddings are received from the central server, and federated aggregation and early alternation are used to achieve highly sensitive and flexible adaptive enhancement of additional local financial entity data and global data. g_rel is used to fuse rel_i to enhance the local model.

9. A federated relationship-aware learning device based on financial knowledge graphs, characterized in that, The device includes: The first module is used to receive local updates and local financial relationship embeddings from the client. The global model is trained in a cross-domain parallel adaptive manner and alternately fuses the embedded relationship-aware federated learning part and the local model during the training process. It uses financial relationship data to perform federated aggregation and alternation to complete the financial knowledge graph completion and distributes the global model weights and global financial relationships to each client. The second module is used to receive global financial relationship embeddings from the central server, use alternating fusion, use local financial entities embedded in ent_i to improve the global model and global financial relationship embeddings, update the financial relationship embeddings through contextual relationships and relationship paths, and send the locally updated weights and financial relationship embeddings into rel_i to the central server. The third module is used to receive global financial relationship embeddings from the central server, and leverages federated aggregation and early alternation to achieve highly sensitive and flexible adaptive enhancement of additional local financial entity data and global data. It uses g_rel to fuse rel_i to enhance the local model.

10. A federated relationship-aware learning device based on financial knowledge graphs, characterized in that, The device includes: At least one processor; At least one memory for storing at least one program; When at least one of the programs is executed by at least one of the processors, it implements the federated relationship-aware learning system based on financial knowledge graphs as described in any one of claims 1-7 or the federated relationship-aware learning method based on financial knowledge graphs as described in claim 8.

Images